人工神经网络在遥感影像分类中的应用与对比研究
|
摘要
随着遥感数据的空间分辨率、光谱分辨率、时间分辨率不断提高,现已形成集多源、多光谱、多空间分辨率的遥感影像海量数据集。但如何利用这些数据来挖掘潜藏的信息,是限制遥感发展和应用的一个关键性问题。其中,影像分类作为遥感信息提取的一个重要方面,应用传统贝叶斯分类器的遥感影像分类方法已不能满足精度的需要,而基于非线性映射的人工神经网络分类方法为这一问题提供了更加理想的解决方案,因为神经网络分类并不是基于某个假定的概率分布,而是通过对训练样本的学习获得网络的权值,形成分类器。采用神经网络算法进行遥感影像分类,可以在一定程度上消除传统的遥感影像分类所带来的模糊性和不确定性。
本文采用2002年呼和浩特东北部地区的Landsat TM遥感数据为数据源,在深入研究标准BP神经网络理论的基础上,通过对训练样本构建方法、训练算法的选择、最佳隐含层神经元个数的确定等方面,对BP神经网络分类过程进行了系统的研究,对于实现过程中所存在的问题,提出自己的改进方法。研究结果表明:使用样本均值法构建的训练样本所训练的网络即使能够以高精度、快速收敛,但不能实现有效分类;针对于标准BP网络收敛速度慢、容易陷入局部最小的问题,研究中发现基于LM算法BP神经网络极大地提高训练速度以及收敛精度,取得非常好的分类效果;在以上研究的基础上,为了克服波段间高相关和数据冗余的问题,将TM1、TM2、TM3共3个波段进行主成分分析,取第一主成分,与其它波段构成数据集,并用此数据进行人工神经网络训练和分类研究。结果表明,在分类精度相差不大的条件下,应用主成分变换使网络的收敛更加迅速、仿真速度有所加快。
将基于LM训练算法的BP神经网络分类结果分别同最大似然法、目视解译结果进行比对,总体上说,神经网络的分类精度要比最大似然法分类效果要好,更加接近目视解译结果,人工神经网络遥感影像分类方法比传统分类方法有所改进。然而,采用神经网络模型仍存在许多亟待解决的问题,如神经网络最佳结构的确定,参数的选择,同时神经网络也是基于光谱特征对遥感影像进行分类的,也存在同最大似然法同样的问题,不能解决“同谱异物”“同物异谱”现象等问题,有待集成其它数据来解决,进一步提高分类精度,满足实际应用的需要。
With remote sensing data spatial resolution, spectral resolution, the temporal resolution of constant improvement,now it has set a space massive data for multi-source, multi-spectral, multi-resolution remote sensing images. But how to use these data to tap the hidden information is the key issues to restrict the development and application of remote sensing. The image classification method as an important aspect in remote sensing extracting information,using the application of traditional Bayesian classifier,has been unable to meet the needs of accuracy, but artificial neural network classification methods based on non-linear mapping will provide more ideal solution for this issue,beacuse this classification method is not based on an assumptive probability distribution, but through training samples to learn network weights,and make up a classifier. A neural network algorithm for image classification, to a certain extent, can eliminate the ambiguity and uncertainty caused by traditional image classification.
Take the TM remote sensing data in 2002 in Northeast Hohhot as data source, based on in-depth study of standard BP neural network theory, through choosing the training sample building methods and training algorithm, confirming the number of best hidden layer neurons,to system study on the processes of BP neural network classification and to show the improve methods for the problems that exist in the realization process. The results show: if used the training samples constructed with sample mean method,the precision high and the convergence rapid, but it can not be achieved effectively classification; the LM algorithm BP Neural networks ,which is found in the issues that standard BP network has slow convergence and is vulnerable to the smallest of local,greatly improve training speed and accuracy of convergence,and made a very good classification results; basis on the above, in order to overcome the high correlation between the band and data redundancy, will do the principal component analysis on TM1, TM2, TM3 of three band, take the first principal component, set a dataset with other band, and using this data to complete network training and study of classification. The results show that under the similar classification accuracy conditions, the principal component transformation can make more rapid convergence and faster simulation speed.
Comparison the BP neural network classification results based on LM algorithm with the maximum likelihood classification and visual interpretative results, overall, the BP network method is better effects than the maximum likelihood, and closer visual interpretative results. BP neural network classification of remote sensing images is more improved than traditional classification methods .However, using neural network model still has many problems to be solved. For example, determine of the neural networks best structure; the selection of parameters;as neural network also classify on remote sensing images based on spectral characteristics,so it exists the same problem with the maximum likelihood method, does not solve the "different body with same spectrum," "same body with different spectrum", other data to be holp to address and to further improve classification accuracy to meet the practical application needs.
本文采用2002年呼和浩特东北部地区的Landsat TM遥感数据为数据源,在深入研究标准BP神经网络理论的基础上,通过对训练样本构建方法、训练算法的选择、最佳隐含层神经元个数的确定等方面,对BP神经网络分类过程进行了系统的研究,对于实现过程中所存在的问题,提出自己的改进方法。研究结果表明:使用样本均值法构建的训练样本所训练的网络即使能够以高精度、快速收敛,但不能实现有效分类;针对于标准BP网络收敛速度慢、容易陷入局部最小的问题,研究中发现基于LM算法BP神经网络极大地提高训练速度以及收敛精度,取得非常好的分类效果;在以上研究的基础上,为了克服波段间高相关和数据冗余的问题,将TM1、TM2、TM3共3个波段进行主成分分析,取第一主成分,与其它波段构成数据集,并用此数据进行人工神经网络训练和分类研究。结果表明,在分类精度相差不大的条件下,应用主成分变换使网络的收敛更加迅速、仿真速度有所加快。
将基于LM训练算法的BP神经网络分类结果分别同最大似然法、目视解译结果进行比对,总体上说,神经网络的分类精度要比最大似然法分类效果要好,更加接近目视解译结果,人工神经网络遥感影像分类方法比传统分类方法有所改进。然而,采用神经网络模型仍存在许多亟待解决的问题,如神经网络最佳结构的确定,参数的选择,同时神经网络也是基于光谱特征对遥感影像进行分类的,也存在同最大似然法同样的问题,不能解决“同谱异物”“同物异谱”现象等问题,有待集成其它数据来解决,进一步提高分类精度,满足实际应用的需要。
With remote sensing data spatial resolution, spectral resolution, the temporal resolution of constant improvement,now it has set a space massive data for multi-source, multi-spectral, multi-resolution remote sensing images. But how to use these data to tap the hidden information is the key issues to restrict the development and application of remote sensing. The image classification method as an important aspect in remote sensing extracting information,using the application of traditional Bayesian classifier,has been unable to meet the needs of accuracy, but artificial neural network classification methods based on non-linear mapping will provide more ideal solution for this issue,beacuse this classification method is not based on an assumptive probability distribution, but through training samples to learn network weights,and make up a classifier. A neural network algorithm for image classification, to a certain extent, can eliminate the ambiguity and uncertainty caused by traditional image classification.
Take the TM remote sensing data in 2002 in Northeast Hohhot as data source, based on in-depth study of standard BP neural network theory, through choosing the training sample building methods and training algorithm, confirming the number of best hidden layer neurons,to system study on the processes of BP neural network classification and to show the improve methods for the problems that exist in the realization process. The results show: if used the training samples constructed with sample mean method,the precision high and the convergence rapid, but it can not be achieved effectively classification; the LM algorithm BP Neural networks ,which is found in the issues that standard BP network has slow convergence and is vulnerable to the smallest of local,greatly improve training speed and accuracy of convergence,and made a very good classification results; basis on the above, in order to overcome the high correlation between the band and data redundancy, will do the principal component analysis on TM1, TM2, TM3 of three band, take the first principal component, set a dataset with other band, and using this data to complete network training and study of classification. The results show that under the similar classification accuracy conditions, the principal component transformation can make more rapid convergence and faster simulation speed.
Comparison the BP neural network classification results based on LM algorithm with the maximum likelihood classification and visual interpretative results, overall, the BP network method is better effects than the maximum likelihood, and closer visual interpretative results. BP neural network classification of remote sensing images is more improved than traditional classification methods .However, using neural network model still has many problems to be solved. For example, determine of the neural networks best structure; the selection of parameters;as neural network also classify on remote sensing images based on spectral characteristics,so it exists the same problem with the maximum likelihood method, does not solve the "different body with same spectrum," "same body with different spectrum", other data to be holp to address and to further improve classification accuracy to meet the practical application needs.
引文
[1] 罗扬帆,基于 BP 神经网络的遥感影像分类研究[D],北京林业大学学位论文,2007
[2] 许仲林,基于改进算法的 BP 神经网络在遥感影像中的应用研究[D],兰州大学硕士学位论文,2007.5
[3] 张治国,人工神经网络及其在地学研究中的应用[D],吉林大学博士学位论文,2006
[4] 杨存建,魏一鸣,陈德清等,基于遥感的洪水灾害承灾体神经网络的提取方法探讨[J],灾害学,1998,13(4): 1～6
[5] 贾永红,人工神经网络在多源遥感影像分类中的应用[J],测绘通报,2000.7:7～8
[6] 罗小波等,RBF 神经网络在遥感影像分类中的应用研究[J],遥感技术与应用,2004
[7] 哈斯巴干,神经网络及其组合算法的遥感数据分类研究[D],中国科学院研究生院博士学位论文,2005
[8] 骆剑承,周成虎等,基于径向基函数(RBF)映射理论的遥感影像分类模型研究[J],中国图像图形学报,2000,5(2)
[9] 闻新,周露等,MATLAB 神经网络仿真与应用[M],科学出版社,2003.7
[10] 贾永红,张春森,王爱平基于神经网络的多源遥感影像分类[J],西安科技学院学报,2001,3(1):58-60
[11] 飞思科技产品研发中心, MATLAB6.5 辅助神经网络分析与设计[M],北京:电子工业出版社,2003
[12] 高隽,人工神经网络原理及仿真实例[M](第 2 版),北京:机械工业出版社,2007.2
[13] 曾志远,卫星遥感图像计算机分类与地学应用研究[M],北京:科学出版社,2004,
[14] MATLAB7.1 帮助文件
[15] 桂现才,BP 神经网络在 MATLAB 上的实现与应用[J],湛江师范学院学报,2004.25(3)4
[16] 丛爽,面向 MATLAB 工具箱的神经网络理论与应用[M],合肥:中国科技大学出版社 1998
[17] 周开利,康耀红,神经网络模型及其 MATLAB 仿真程序设计[M],清华大学出版社,2005
[18] 袁曾任,人工神经元网络及其应用[M],北京清华大学出版社,1999
[19] 王双亭,朱宝山编著,遥感图像判绘[M],解放军出版社,2002
[20] 梅安新,彭望碌,秦其明,刘慧平,遥感导论[M],高等教育出版社,2001
[21] 梅长林,范金城,数据分析方法[M],高等教育出版社,2006
[22] 周成虎,骆剑承,杨小梅等编著,遥感影像地学理解与分析[M],科学出版社,2001
[23]于凤鸣,基于人工神经网络的遥感影像模式分类研究[J],内蒙古师范大学学报,2007
[24] 马建文,李启青,哈斯巴干,戴芹等著,遥感数据智能处理方法与程序设计[M],科学出版社,2005
[25] 赵红,人工神经万网络在高分辨率遥感影像森林植被分类中的应用研究[D],西安科技大学硕士学位论文,2006.4
[26] 张志涌,精通 MATLAB6.5 版[M],北京航空航天大学出版社,2007.8
[27] [美] D. Hanselman , B ,Littlefield 著,张航,黄攀译,精通 MATLAB6 A Comprehensive Tutorial and Reference 清华大学出版社,2002.6
[28] 任春涛,基于遥感监测的湖泊富营养化状态的模式识别研究[D],内蒙古农业大学博士学位论文,2007.4
[29] 刘玉洁,杨忠东等,MODIS 遥感信息处理原理与算法[M],北京:科学出版社,2001
[30] 陈述彭,赵英时,遥感地学分析[M],北京:测绘出版社,1990
[31] 刘伟,基于人工神经网络的遥感影像模式分类研究[D],沈阳工业大学硕士学位论文,2005
[2] 许仲林,基于改进算法的 BP 神经网络在遥感影像中的应用研究[D],兰州大学硕士学位论文,2007.5
[3] 张治国,人工神经网络及其在地学研究中的应用[D],吉林大学博士学位论文,2006
[4] 杨存建,魏一鸣,陈德清等,基于遥感的洪水灾害承灾体神经网络的提取方法探讨[J],灾害学,1998,13(4): 1～6
[5] 贾永红,人工神经网络在多源遥感影像分类中的应用[J],测绘通报,2000.7:7～8
[6] 罗小波等,RBF 神经网络在遥感影像分类中的应用研究[J],遥感技术与应用,2004
[7] 哈斯巴干,神经网络及其组合算法的遥感数据分类研究[D],中国科学院研究生院博士学位论文,2005
[8] 骆剑承,周成虎等,基于径向基函数(RBF)映射理论的遥感影像分类模型研究[J],中国图像图形学报,2000,5(2)
[9] 闻新,周露等,MATLAB 神经网络仿真与应用[M],科学出版社,2003.7
[10] 贾永红,张春森,王爱平基于神经网络的多源遥感影像分类[J],西安科技学院学报,2001,3(1):58-60
[11] 飞思科技产品研发中心, MATLAB6.5 辅助神经网络分析与设计[M],北京:电子工业出版社,2003
[12] 高隽,人工神经网络原理及仿真实例[M](第 2 版),北京:机械工业出版社,2007.2
[13] 曾志远,卫星遥感图像计算机分类与地学应用研究[M],北京:科学出版社,2004,
[14] MATLAB7.1 帮助文件
[15] 桂现才,BP 神经网络在 MATLAB 上的实现与应用[J],湛江师范学院学报,2004.25(3)4
[16] 丛爽,面向 MATLAB 工具箱的神经网络理论与应用[M],合肥:中国科技大学出版社 1998
[17] 周开利,康耀红,神经网络模型及其 MATLAB 仿真程序设计[M],清华大学出版社,2005
[18] 袁曾任,人工神经元网络及其应用[M],北京清华大学出版社,1999
[19] 王双亭,朱宝山编著,遥感图像判绘[M],解放军出版社,2002
[20] 梅安新,彭望碌,秦其明,刘慧平,遥感导论[M],高等教育出版社,2001
[21] 梅长林,范金城,数据分析方法[M],高等教育出版社,2006
[22] 周成虎,骆剑承,杨小梅等编著,遥感影像地学理解与分析[M],科学出版社,2001
[23]于凤鸣,基于人工神经网络的遥感影像模式分类研究[J],内蒙古师范大学学报,2007
[24] 马建文,李启青,哈斯巴干,戴芹等著,遥感数据智能处理方法与程序设计[M],科学出版社,2005
[25] 赵红,人工神经万网络在高分辨率遥感影像森林植被分类中的应用研究[D],西安科技大学硕士学位论文,2006.4
[26] 张志涌,精通 MATLAB6.5 版[M],北京航空航天大学出版社,2007.8
[27] [美] D. Hanselman , B ,Littlefield 著,张航,黄攀译,精通 MATLAB6 A Comprehensive Tutorial and Reference 清华大学出版社,2002.6
[28] 任春涛,基于遥感监测的湖泊富营养化状态的模式识别研究[D],内蒙古农业大学博士学位论文,2007.4
[29] 刘玉洁,杨忠东等,MODIS 遥感信息处理原理与算法[M],北京:科学出版社,2001
[30] 陈述彭,赵英时,遥感地学分析[M],北京:测绘出版社,1990
[31] 刘伟,基于人工神经网络的遥感影像模式分类研究[D],沈阳工业大学硕士学位论文,2005